Retrieval by Classification: Discriminative Binary Embedding for Sketch-Based Image Retrieval

Abstract

Sketch-based image retrieval (SBIR) intends to use free-hand sketch drawings as query to retrieve correlated real-world images from database. Hashing based methods gradually become the mainstream approaches in SBIR with its low memory usage and high query speed. Existing hashing based methods are incapable of guiding hash codes to preserve inter-class relationship and improving object recognition ability of hash functions simultaneously, which limits the higher performance. Hence, we propose Discriminative Binary Embedding (DBE), a novel algorithm of considering inter-class relationship and object recognition ability in a joint manner by treating retrieval as classification. Specifically, we apply NLP methods to encode category labels as binary embedding and then build classifiers for images and sketches, so as to obtain hash codes of instances based on binary embedding of predicted labels. Experimental results on two benchmarks show that DBE outperforms several state-of-the-arts.